Conventionally, a moving object detection apparatus is disclosed in Japanese Patent Laid-Open No. 2003-288600 (JP '600). The moving object detection apparatus of JP '600 detects a pedestrian in a camera capture image, by extracting a motion vector from a sampled capture image, by applying a parametric eigenspace method to the motion vector, and by generating a reference curve data. The reference curve data is based on a projection of the motion vector to an eigenspace while extracting a motion vector from the capture image that is considered as an object of pedestrian extraction. The apparatus of JP '600 further applies the parametric eigenspace method to the motion vector, and gets information on projection points of the motion vector to the eigenspace derived from the eigenspace method. That is, by comparing the projection points with the reference curve, the pedestrian in the capture image is detected.
Further, in another method is disclosed in Japanese Patent Laid-Open No. 2005-123968 (JP '968). The apparatus of JP '968 calculates a motion vector of a characteristic point from a capture image, and detects an approaching moving object that is approaching a subject vehicle as a group of characteristic points. The apparatus of JP '968 emphasizes an image area including such moving object in the capture image. Also, in Japanese Patent Laid-Open No. 2005-132170 (JP '170), the apparatus detects a moving object as a group of substantially same size motion vectors based on the motion vector of the characteristic point.
Further, a Harris method and a Karhunen-Loeve Transform (KLT) method are known to extract a characteristic point from the capture image. According to the KLT method, it is possible to extract a motion vector (i.e., an optical flow) of the characteristic point by tracking the extracted characteristic point.
Furthermore, as for the calculation method of the motion vector, a gradient method and a block matching method are known as well, and, as an application of these techniques, Japanese Patent Laid-Open No. 2007-172540 (JP '540) discloses a calculation of a motion vector of image pixels in an image area that correspond to a far field in the capture image based on the gradient method and a calculation of a motion vector of image pixels in an image area that correspond to a near field in the capture image.
Further, a combination of plural methods for detecting a moving object is disclosed in Japanese Patent Laid-Open No. 2009-146153 (JP '153). The method in JP '153 detects a moving object based on the motion vector calculated by the block matching method when the vehicle is not stopping, and detects a moving object based on the differential capture image of two frames.
Further, a moving object is detected by detecting a vanishing point where the moving object vanishes from the screen and by analyzing a surrounding of the vanishing point in a known technique disclosed in a Japanese Patent Laid-Open No. 2009-143385 (JP '385).
In the conventional techniques, a characteristic point and a motion vector of the far-side approaching vehicle can be detected in a strictly limited case, that is, requiring at least 5 frame intervals (i.e., more than 166 ms) to detect the motion vector of the far-side approaching vehicle. In addition, even when the motion vector is detected, it may be difficult, in many cases, to distinguish the motion vector from the background noise.
Further, since the amount of calculation for detecting the motion vector is large, implementation of such calculation by using simple (i.e., “light”) software has been considered as difficult.